图神经网络参与湍流数值计算的两条路线:直接推进与辅助粗网格求解的能力、未解问题与研究方向比较
湍流物理基础与数据基础设施:HIT 统计理论、能谱/能通量 DNS 证据与无散度合成湍流生成
这组文献不涉及任何学习模型,而是两条路线共同的物理标尺与数据入口:均匀各向同性湍流的经典统计理论(K41、能谱与耗散)、近耗散区能谱与能通量关系的高保真伪谱 DNS 证据,以及按指定能谱生成无散度随机场和散度自由合成涡入口的构造方法。它们界定了报告中必须核查的量(短时精度、能谱、能量耗散、散度残差),也说明两条路线对 DNS 轨迹与入口/边界条件数据的依赖程度与获取成本,是判断 GNN 是否守住物理统计的前置参照。
- Physics of Homogeneous Isotropic Turbulence(John Newman, J. Trainham, 2024, Russian journal of electrochemistry)
- Kinetic energy spectrum and transfer within the near-dissipation range of compressible isotropic turbulence(Guo-Wei Yang, Zheng Yan, Jun-Feng Wu, Jianchun Wang, Chang-Ping Yu, Xin-Liang Li, Li-Feng Wang, 2025, The Physics of Fluids)
- High-fidelity pseudo-spectral simulations of isotropic turbulence: Large eddy simulations with semi-adaptive parametrization of residual flow scales(Michał Rajek, J. Pozorski, 2026, The Physics of Fluids)
- A novel vector potential random flow generation method for synthesizing divergence-free homogeneous isotropic turbulence with arbitrary spectra(Chao Li, Lingwei Chen, Jing-Han Wang, Wentong Zhang, Xiang-Jie Wang, Zhuoran Wang, Gang Hu, 2024, The Physics of Fluids)
- A new divergence-free synthetic eddy method for generating homogeneous isotropic turbulence with a prescribed energy spectrum(Yun-Zhu Cai, Jia-Wei Wan, A. Kareem, 2023, Computers & Fluids)
问题诊断与可学性边界:先验—后验鸿沟、领域综述与粗网格修正的负面结果
这组提供批判性理论标尺与明确的负面证据:一篇综述把 RANS 病态、a priori/a posteriori 差距与 PINN 谱偏差归因于'逐点离线训练目标'与'求解器耦合多尺度统计表现'的结构性错配;一篇 GNN 领域综述梳理流场预测、湍流建模与复杂流动仿真中的架构、网格适配、训练成本与泛化瓶颈;一篇在非结构四面体网格、转捩流动上给出前馈网络、多种 GNN 与层次网络均无法给出可用粗网格修正的负面结果。三者共同支撑报告对'当前空缺在精度、稳定性、物理统计、跨工况还是公平对照'的定位判断,并警示应优先选择能隔离 GNN 贡献的受控算例。
- Bridging physics and data: A review of machine learning approaches for turbulence closure modeling(Jia-Hao Hu, Jun-Ya Yang, 2026, The Physics of Fluids)
- Application of Graph Neural Networks in Computational Fluid Mechanics(An-Yi Xiao, 2026, Applied and Computational Engineering)
- On Limitations of Several Approaches for Correcting Coarse Flow Simulations With Machine Learning Techniques(Sarah Katz, V. John, 2026, ZAMM ‐ Journal of Applied Mathematics and Mechanics / Zeitschrift für Angewandte Mathematik und Mechanik)
GNN 方法学底座:消息传递基础模型、各向异性与多尺度物理约束表达及可扩展训练架构
这组聚焦两条路线共享的架构与成本基线而非具体湍流路线:原始 GNN 模型给出图域消息传递与监督学习范式;各向异性 GNN 与多尺度物理约束图网络针对湍流强非平衡、尺度间交互与非局部依赖改进表达能力;多处理器并行 PIGNN 与混合精度训练直接关系到算力预算、训练成本与推理效率。它们为后续'相近网格、时间步与计算预算下的公平对照'提供方法学与成本参照。
- The Graph Neural Network Model(F. Scarselli, M. Gori, A. Tsoi, M. Hagenbuchner, G. Monfardini, 2009, IEEE Transactions on Neural Networks)
- The anisotropic graph neural network model with multiscale and nonlinear characteristic for turbulence simulation(Qiang Liu, Wei Zhu, Xi-Yu Jia, Feng Ma, Jun Wen, Yixiong Wu, Kuang-Qi Chen, Zhenhai Zhang, Shuang Wang, 2024, Computer Methods in Applied Mechanics and Engineering)
- Multi-scale graph neural network for physics-informed fluid simulation(Lan Wei, N. Freris, 2024, The Visual Computer)
- Towards Accelerating Physics Informed Graph Neural Network for Fluid Simulation(Yi-Di Wang, F. Guan, Malcolm Yoke Hean Low, Zheng-Kui Wang, Aik Beng Ng, Simon See, 2024, SIGGRAPH Asia 2024 Posters)
路线一(自由滚动):GNN 直接推进的自回归滚动、误差累积抑制与长期稳定性
这组文献共同把流场演化本身交给图网络并递归多步推进,核心矛盾是自回归误差累积导致的长期失稳:图注意力模型在二维圆柱绕流上的滚动对比 GNN 与图神经 ODE 训练模式的稳定性差异;气动预测工作用'两步训练+自回归微调'检验跨声速抖振与流固耦合;ALE 一致的图神经算子 Transformer 修正移动边界处理;物理约束课程学习完成数百步滚动;网格图学习与圆柱入水长期预测则给出 RMSE 随步数增长、R²、速度散度与 VOF 残差的量化曲线,并检验对未见 Froude 数、初始角度、几何与网格拓扑的迁移。它们构成'一步预测 vs 自由滚动推进'区分与长期稳定性、计算成本判断的主要证据。
- Graph attention network-based fluid simulation model(Qiang Liu, Wei Zhu, Feng Ma, Xi-Yu Jia, Yu Gao, Jun Wen, 2022, AIP Advances)
- Spatio-Temporal Prediction of Unsteady Airfoil Aerodynamics Using Augmented Graph Neural Ordinary Differential Equations with Exogenous Controls(H. Lange, Reik Thormann, P. Bekemeyer, 2026, arXiv.org)
- An ALE-Consistent Graph Neural Operator-Transformer Framework for Fluid-Structure Interaction(Shi-Hang Zhao, Martín Saravia, Hao-Kui Jiang, Zhi-Yang Xue, Shun-Xiang Cao, 2026, arXiv.org)
- Physics-Informed Graph Neural Network Surrogates for Turbulent Nanoparticle Dispersion in Dental Clinical Environments(Takshak Shende, Viktor Popov, 2026, arXiv.org)
- Mesh-based graph learning for full-field prediction of stratified circular-cylinder wakes(Tao Zhao, Li Zou, Xue-Yu Wang, Guo-Qing Jin, Zong-Bing Yu, 2026, The Physics of Fluids)
- Long-term prediction of cylinder water entry flow fields using a physics-informed spatiotemporal graph neural network(Chongbin Shi, Gui-Yong Zhang, Teizhi Sun, Yu-Tong Sui, 2026, The Physics of Fluids)
路线一(短时与条件化):单步/短时全场预测精度、高阶结构函数与间歇性统计诊断及零微调泛化
这组属于 GNN 直接推进中评价重点不同的另一侧:不追求长时稳定,而是检验预测场在湍流统计意义上的保真度与条件化泛化能力——用二阶/三阶结构函数核对 Kolmogorov 2/3 与 4/5 律、用扩展自相似标度指数刻画近尾迹各向异性与间歇性;在三维瞬态流动(钝体绕流、动脉瘤血流)与 GNN-CNN 混合架构中检验对未见入流条件与几何的零微调泛化;并以物理约束损失保证质量守恒。它们说明短时精度已可验证,但湍流统计一致性与三维真实场景仍是薄弱环节。
- Spatiotemporal prediction of bluff-body hydrodynamics using the graph neural network(Anjan Samanta, P. Gopmandal, Sankar Sarkar, 2025, The Physics of Fluids)
- Graph and convolutional neural network coupling with a high-performance large-eddy simulation solver(A. Serhani, Victor Xing, D. Dupuy, C. Lapeyre, G. Staffelbach, 2024, Computers & Fluids)
- Physics constrained graph neural network for real time prediction of intracranial aneurysm hemodynamics(V. Lannelongue, Paul Garnier, Pablo Jeken-Rico, A. Goetz, P. Meliga, Y. Chau, E. Hachem, 2026, NPJ Digital Medicine)
路线二(亚格子闭合):GNN 学习亚格子应力与过滤源项并接回 LES 求解器
这组是第二条路线的主力:求解器推进、GNN 只预测闭合项,学习目标明确为亚格子应力或过滤标量生成率(反应源项)等物理量。多尺度自相似输入的 GNN 亚格子应力模型在 HIT、混合层与槽道流上做先验与后验检验并与涡黏模型比较;等变嵌入 GNN 配合对称保持离散与强化学习在 HIT 与槽道中保持可实现性;非均匀网格上的过滤源项闭合用于反应流。它们共同回答'是否真正接回求解器、a posteriori 统计能否恢复'以及'是否与 MLP、传统涡黏在同一滤波宽度与算力预算下对照'。
- Multi-scale graph neural network modeling of near-wall turbulence with robust flow generalizability(L. Hutin, Ye Wang, M. Tanahashi, 2026, The Physics of Fluids)
- Harnessing Equivariance: Modeling Turbulence with Graph Neural Networks(Marius Kurz, Andrea Beck, B. Sanderse, 2025, arXiv.org)
- Modeling subgrid scale production rates on complex meshes using graph neural networks(Priyabrat Dash, M. Bode, K. Aditya, 2026, arXiv.org)
路线二(壁面与非局部应力):壁面剪应力模型、E(3) 等变雷诺应力闭合与求解器在环鲁棒训练
这组把封闭项扩展到壁湍流与 RANS 场景,重点是非局部依赖与数值稳定性:非结构网格 GNN 壁面剪应力模型在槽道、后台阶、叶栅上做 a posteriori 验证(壁面模型路线);E(3) 等变 GNN 结合可实现性投影与求解器在环微调处理 RANS 非局部雷诺应力与病态迭代;鲁棒机器学习训练策略与 solver-aware 闭合以离散残差/数据同化式目标提升耦合后的稳定性。共同点是考察壁湍流与更高雷诺数扩展场景下的泛化与稳定,而非单纯一步预测精度。
- Modeling the wall shear stress in large-eddy simulation using graph neural networks(D. Dupuy, N. Odier, C. Lapeyre, D. Papadogiannis, 2023, Data-Centric Engineering)
- Realisable E(3)-equivariant graph networks with near-wall reinforcement learning for turbulence closure(Fredrik Dubay Myklebust, Y. Li, 2026, Transport Phenomena)
- Exploration of robust machine learning strategy for subgrid scale stress modelling(Asahi Abekawa, Y. Minamoto, K. Osawa, Haruya Shimamoto, M. Tanahashi, 2023, The Physics of Fluids)
- Deep Learning of Solver-Aware Turbulence Closures from Nudged LES Dynamics(Ashwin Suriyanarayanan, Dibyajyoti Chakraborty, Romit Maulik, 2026, arXiv.org)
混合求解路线:粗网格离散误差与残差修正、非结构网格小尺度重构而非整场演化预测
这组既不做整场时间推进,也不严格只学亚格子应力,而是让粗网格 CFD 继续推进、由学习模块修正离散误差、粗—细网格残差或重构小尺度结构。它们直接回答两条路线的成本—精度权衡与网格分辨率边界,并揭示工业级算例无法做系统性网格收敛研究的现实约束,是评估'算力受限下如何量化并降低粗网格误差'的补充路线。
- Correcting the discretization error of coarse grid CFD simulations with machine learning(A. Kiener, S. Langer, P. Bekemeyer, 2022, 8th European Congress on Computational Methods in Applied Sciences and Engineering)
- Diff-KNN: Residual Correction of Baseline Wind Predictions in Urban Settings(D. Nowak, Jennifer Werner, F. Hunger, Tomas Johnson, A. Mark, R. Mitkov, F. Edelvik, 2025, Machine Learning and Knowledge Extraction)
- Super-resolution of turbulent reacting flows on complex meshes using graph neural networks(Priyabrat Dash, K. Aditya, C. Frouzakis, M. Bode, 2026, arXiv.org)
无学习与传统解析闭合基线:应力分解、多分形 SGS、谱涡粘/反卷积与 Volterra 自主闭合
这组代表完全或几乎不依赖机器学习的建模基线,是判断 GNN 增益的物理起点:Leonard/Cross/Reynolds 应力分解与能量—拟涡能谱约束下的 LES 先验/后验分析、多分形 SGS 模型、自适应谱涡粘与直接反卷积模型,以及以 Volterra 级数在线反演辨识的自主闭合。它们提供可解释、可复现的闭合性能参照,并示范了 a priori 与 a posteriori 双检验的成熟范式。
- A priori and a posteriori analysis in Large eddy simulation of the two-dimensional decaying turbulence using Explicit filtering approach(P. Perezhogin, A. Glazunov, 2021, No journal)
- Multifractal subgrid-scale modeling for large-eddy simulation. II. Backscatter limiting and a posteriori evaluation(G. Burton, W. Dahm, 2005, Physics of Fluids)
- Advances in direct deconvolution modeling of subgrid-scales for flows with discontinuities(N. Adams, 2000, No journal)
- A NEW AUTONOMIC CLOSURE FOR LARGE EDDY SIMULATIONS(Ryan N. King, P. Hamlington, W. Dahm, 2015, Proceeding of Ninth International Symposium on Turbulence and Shear Flow Phenomena)
非图机器学习闭合基线(一):ANN/FCNN 代数与梯度模型在可压缩 HIT 上的先验—后验对照
这组是全连接/局部模板神经网络亚格子模型的经典对照系,在相近网格、滤波宽度与算力下与原位 GNN 模型构成最直接的公平比较:涵盖可压缩/不可压均匀各向同性湍流中的非线性代数模型与空间梯度模型、含测试滤波输入与融合策略的设计,以及跨网格分辨率、跨马赫数与跨雷诺数的外推实验。它们用于隔离'图结构'本身带来的增益。
- Artificial-neural-network-based nonlinear algebraic models for large-eddy simulation of compressible wall-bounded turbulence(Dehao Xu, Jianchun Wang, Chang-Ping Yu, Shi-Yi Chen, 2023, Journal of Fluid Mechanics)
- Artificial neural network-based spatial gradient models for large-eddy simulation of turbulence(Yun-Peng Wang, Ze-Long Yuan, Chen-Yue Xie, Jianchun Wang, 2021, AIP Advances)
- Subgrid-scale model for large-eddy simulation of isotropic turbulent flows using an artificial neural network(Z. Zhou, G. He, Shi-Zhao Wang, G. Jin, 2019, Computers & Fluids)
- Artificial neural network approach to large-eddy simulation of compressible isotropic turbulence.(Chen-Yue Xie, Jianchun Wang, Ke Li, Chao Ma, 2019, Physical Review E)
非图机器学习闭合基线(二):混合模型、NN-LES 工程化落地与钝体绕流等扩展场景应用
这组把非图神经网络闭合推向工程实现与复杂几何:ANN 混合模型组合显式与隐式/涡黏成分以改善后向散射与稳定性,NN 驱动的 LES 工程化流程检验计算开销与实际收益,钝体绕流的 LES 应用则检验在分离、剪切层主导流动中的表现。它们既是 GNN 路线在壁面/绕流等扩展场景的直接对照,也说明基线在跨工况外推上的既有局限。
- Artificial neural network mixed model for large eddy simulation of compressible isotropic turbulence(Chen-Yue Xie, Jianchun Wang, Hui Li, M. Wan, Shi-Yi Chen, 2019, The Physics of Fluids)
- Toward neural-network-based large eddy simulation: application to turbulent channel flow(Jonghwa Park, H. Choi, 2021, Journal of Fluid Mechanics)
- Large eddy simulation of flow over a circular cylinder with a neural-network-based subgrid-scale model(Myunghwa Kim, Jonghwan Park, H. Choi, 2024, Journal of Fluid Mechanics)
合并去重后,42 篇文献被组织为 11 个并列分组。分组结构呈"共同基础—问题诊断—方法学底座—路线一(两组)—路线二(两组)—混合求解—三类对照基线"的逻辑:第 1 组给出 HIT 统计理论、能谱/能通量 DNS 证据与无散度合成湍流入口等物理标尺与数据入口,界定"短时精度、能谱、能量耗散、散度残差"应被核查的量;第 2 组用综述与负面结果定位先验—后验鸿沟与粗网格修正的可学性边界;第 3 组汇总 GNN 架构与可扩展训练成本基线。第 4、5 组按评价重点区分路线一内部的两类能力——自由滚动多步推进的长期稳定性与误差累积抑制,以及单步/短时与条件化预测的精度、高阶结构函数与间歇性统计诊断。第 6、7 组对应路线二的不同学习目标与物理含义:亚格子应力/过滤源项接回 LES 求解器,以及壁面剪应力、E(3) 等变雷诺应力与求解器在环鲁棒训练;第 8 组为粗网格离散误差/残差修正与小尺度重构的混合路线。第 9、10、11 组刻意把"非图机器学习闭合"与"无学习解析闭合"从 GNN 方法中单独拆出,构成同网格、同滤波宽度、同算力预算下隔离 GNN 贡献的公平对照来源。总体判断:周期域均匀各向同性湍流仍是训练与先验检验的主场景,壁湍流、绕流、叶栅、反应流等作为扩展场景;当前空缺更多集中在长期稳定性与物理统计保真、跨工况外推,以及缺乏时间步/网格/算力对齐的公平对照,而非单纯一步预测精度。
这组文献不涉及任何学习模型,而是两条路线共同的物理标尺与数据入口:均匀各向同性湍流的经典统计理论(K41、能谱与耗散)、近耗散区能谱与能通量关系的高保真伪谱 DNS 证据,以及按指定能谱生成无散度随机场和散度自由合成涡入口的构造方法。它们界定了报告中必须核查的量(短时精度、能谱、能量耗散、散度残差),也说明两条路线对 DNS 轨迹与入口/边界条件数据的依赖程度与获取成本,是判断 GNN 是否守住物理统计的前置参照。相关文献:John Newman et. al, 2024 等 5 篇文献
这组提供批判性理论标尺与明确的负面证据:一篇综述把 RANS 病态、a priori/a posteriori 差距与 PINN 谱偏差归因于'逐点离线训练目标'与'求解器耦合多尺度统计表现'的结构性错配;一篇 GNN 领域综述梳理流场预测、湍流建模与复杂流动仿真中的架构、网格适配、训练成本与泛化瓶颈;一篇在非结构四面体网格、转捩流动上给出前馈网络、多种 GNN 与层次网络均无法给出可用粗网格修正的负面结果。三者共同支撑报告对'当前空缺在精度、稳定性、物理统计、跨工况还是公平对照'的定位判断,并警示应优先选择能隔离 GNN 贡献的受控算例。相关文献:Jia-Hao Hu et. al, 2026 等 3 篇文献
这组聚焦两条路线共享的架构与成本基线而非具体湍流路线:原始 GNN 模型给出图域消息传递与监督学习范式;各向异性 GNN 与多尺度物理约束图网络针对湍流强非平衡、尺度间交互与非局部依赖改进表达能力;多处理器并行 PIGNN 与混合精度训练直接关系到算力预算、训练成本与推理效率。它们为后续'相近网格、时间步与计算预算下的公平对照'提供方法学与成本参照。相关文献:F. Scarselli et. al, 2009 等 4 篇文献
这组文献共同把流场演化本身交给图网络并递归多步推进,核心矛盾是自回归误差累积导致的长期失稳:图注意力模型在二维圆柱绕流上的滚动对比 GNN 与图神经 ODE 训练模式的稳定性差异;气动预测工作用'两步训练+自回归微调'检验跨声速抖振与流固耦合;ALE 一致的图神经算子 Transformer 修正移动边界处理;物理约束课程学习完成数百步滚动;网格图学习与圆柱入水长期预测则给出 RMSE 随步数增长、R²、速度散度与 VOF 残差的量化曲线,并检验对未见 Froude 数、初始角度、几何与网格拓扑的迁移。它们构成'一步预测 vs 自由滚动推进'区分与长期稳定性、计算成本判断的主要证据。相关文献:Qiang Liu et. al, 2022 等 6 篇文献
这组属于 GNN 直接推进中评价重点不同的另一侧:不追求长时稳定,而是检验预测场在湍流统计意义上的保真度与条件化泛化能力——用二阶/三阶结构函数核对 Kolmogorov 2/3 与 4/5 律、用扩展自相似标度指数刻画近尾迹各向异性与间歇性;在三维瞬态流动(钝体绕流、动脉瘤血流)与 GNN-CNN 混合架构中检验对未见入流条件与几何的零微调泛化;并以物理约束损失保证质量守恒。它们说明短时精度已可验证,但湍流统计一致性与三维真实场景仍是薄弱环节。相关文献:Anjan Samanta et. al, 2025 等 3 篇文献
这组是第二条路线的主力:求解器推进、GNN 只预测闭合项,学习目标明确为亚格子应力或过滤标量生成率(反应源项)等物理量。多尺度自相似输入的 GNN 亚格子应力模型在 HIT、混合层与槽道流上做先验与后验检验并与涡黏模型比较;等变嵌入 GNN 配合对称保持离散与强化学习在 HIT 与槽道中保持可实现性;非均匀网格上的过滤源项闭合用于反应流。它们共同回答'是否真正接回求解器、a posteriori 统计能否恢复'以及'是否与 MLP、传统涡黏在同一滤波宽度与算力预算下对照'。相关文献:L. Hutin et. al, 2026 等 3 篇文献
这组把封闭项扩展到壁湍流与 RANS 场景,重点是非局部依赖与数值稳定性:非结构网格 GNN 壁面剪应力模型在槽道、后台阶、叶栅上做 a posteriori 验证(壁面模型路线);E(3) 等变 GNN 结合可实现性投影与求解器在环微调处理 RANS 非局部雷诺应力与病态迭代;鲁棒机器学习训练策略与 solver-aware 闭合以离散残差/数据同化式目标提升耦合后的稳定性。共同点是考察壁湍流与更高雷诺数扩展场景下的泛化与稳定,而非单纯一步预测精度。相关文献:D. Dupuy et. al, 2023 等 4 篇文献
这组既不做整场时间推进,也不严格只学亚格子应力,而是让粗网格 CFD 继续推进、由学习模块修正离散误差、粗—细网格残差或重构小尺度结构。它们直接回答两条路线的成本—精度权衡与网格分辨率边界,并揭示工业级算例无法做系统性网格收敛研究的现实约束,是评估'算力受限下如何量化并降低粗网格误差'的补充路线。相关文献:A. Kiener et. al, 2022 等 3 篇文献
这组代表完全或几乎不依赖机器学习的建模基线,是判断 GNN 增益的物理起点:Leonard/Cross/Reynolds 应力分解与能量—拟涡能谱约束下的 LES 先验/后验分析、多分形 SGS 模型、自适应谱涡粘与直接反卷积模型,以及以 Volterra 级数在线反演辨识的自主闭合。它们提供可解释、可复现的闭合性能参照,并示范了 a priori 与 a posteriori 双检验的成熟范式。相关文献:P. Perezhogin et. al, 2021 等 4 篇文献
这组是全连接/局部模板神经网络亚格子模型的经典对照系,在相近网格、滤波宽度与算力下与原位 GNN 模型构成最直接的公平比较:涵盖可压缩/不可压均匀各向同性湍流中的非线性代数模型与空间梯度模型、含测试滤波输入与融合策略的设计,以及跨网格分辨率、跨马赫数与跨雷诺数的外推实验。它们用于隔离'图结构'本身带来的增益。相关文献:Dehao Xu et. al, 2023 等 4 篇文献
这组把非图神经网络闭合推向工程实现与复杂几何:ANN 混合模型组合显式与隐式/涡黏成分以改善后向散射与稳定性,NN 驱动的 LES 工程化流程检验计算开销与实际收益,钝体绕流的 LES 应用则检验在分离、剪切层主导流动中的表现。它们既是 GNN 路线在壁面/绕流等扩展场景的直接对照,也说明基线在跨工况外推上的既有局限。相关文献:Chen-Yue Xie et. al, 2019 等 3 篇文献
总计 42 篇相关文献
No abstract available
A new graph neural network (GNN) subgrid-scale (SGS) stress modeling framework based on multi-scale self-similarity input features is investigated as an improvement in generalizability over a previous GNN model developed by the authors' group. Training and a priori testing are conducted with an incompressible direct numerical simulation (DNS) database comprising homogeneous isotropic turbulence (HIT), turbulent mixing layer, and turbulent channel flow (CH) cases, each across five Reynolds numbers. DNS fields are filtered at two different scales using both Gaussian and box kernels to create a diverse large eddy simulation training dataset. The multi-scale Leonard stress tensor and spatial velocity fluctuations are used as inputs to capture scale interactions from a broader portion of the turbulent energy spectrum. Galilean invariance and scale equivariance are ensured by input definition and nondimensionalization, respectively. Randomized data augmentation is used to improve model robustness to changes in orientation. Spline-based graph-convolution layers enable non-local feature extraction while maintaining model compatibility with unstructured grids. A priori analysis of four GNN SGS models shows that graph convolution and multi-scale input features provide significant improvement in accuracy for incompressible flows at Reynolds numbers below, within and above the training range, with particular impact in the near-wall region. Model generalizability to compressible reactive flows is high except in a wall-bounded case. In a posteriori testing with incompressible turbulent CH, one multi-scale GNN was found to better match reference DNS profiles than conventional eddy-viscosity models, with computational overhead comparable to that of a dynamic k-equation model.
Computational fluid dynamics (CFD) discretizes the Navier-Stokes equations on computational meshes to obtain flow-field quantities like velocity, pressure, and density through numerical iteration, but traditional solvers, though reliable in accuracy, are computationally expensive and time-consuming for high-Reynolds-number turbulence and complex unsteady flows. With geometric deep learning, graph neural networks (GNNs) overcome the limitation of conventional neural network (CNNs) that only handle grid-structured Euclidean data; by message-passing, they naturally fit unstructured CFD meshes and capture spatial correlations. This paper systematically reviews GNN fundamentals and core fluid mechanics, including turbulence characteristics and mesh type classifications, and then focuses on typical GNN applications in flow-field prediction, turbulence modeling, and complex-flow simulation. Existing challenges such as high training costs, insufficient generalization, and limited computational efficiency are analyzed, and promising future research directions are discussed. The review indicates GNNs are efficient surrogates for traditional solvers, offering a pathway to intelligent, fast, high-fidelity simulation.
Traditional computational fluid dynamics (CFD) techniques deduce the dynamic variations in flow fields by using finite elements or finite differences to solve partial differential equations. CFD usually involves several tens of thousands of grid nodes, which entail long computation times and significant computational resources. Fluid data are usually irregular data, and there will be turbulence in the flow field where the physical quantities between adjacent grid nodes are extremely nonequilibrium. We use a graph attention neural network to build a fluid simulation model (GAFM). GAFM assigns weights to adjacent node-pairs through a graph attention mechanism. In this way, it is not only possible to directly calculate the fluid data but also to adjust for nonequilibrium in vortices, especially turbulent flows. The GAFM deductively predicts the dynamic variations in flow fields by using spatiotemporally continuous sample data. A validation of the proposed GAFM against the two-dimensional (2D) flow around a cylinder confirms its high prediction accuracy. In addition, the GAFM achieves faster computation speeds than traditional CFD solvers by two to three orders of magnitude. The GAFM provides a new idea for the rapid optimization and design of fluid mechanics models and the real-time control of intelligent fluid mechanisms.
Fluid simulations in Computational Fluid Dynamics (CFD) often involve solving complex Partial Differential Equations (PDEs) with varying parameters, leading to high computational costs and slow inference times. Graph neural networks (GNNs) offer a promising alternative by efficiently representing mesh spaces and reducing these costs. In this work, we introduce a pioneering Physics-Informed Graph Neural Network (PIGNN) approach, utilizing a multi-GNN processor architecture that assigns each output parameter to a dedicated GNN processor with individual decoders. This innovative method reduces the training epochs for PIGNN by half while preserving accuracy. Additionally, we adopted mixed-precision training and achieved further improvements in reducing the training time to one-quarter of the original network while maintaining VRAM usage.
No abstract available
Various aspects of machine learning (ML) are explored to resolve limitations appearing in current ML-based subgrid scale (SGS) stress modelling. Graph neural network (GNN), applied in the present study, allows flexible and rigorous use of spatial convolution regardless of the proximity to physical boundaries and mesh uniformity. Along with GNN, the proposed feature scaling method relies only on the local quantities and can be applied for a range of flow configurations. A data augmentation method is also proposed to consider the rotational invariant. All these techniques are implemented in the present model, and the model is compared with versions of corresponding ML-based models including a typical multilayer perceptron (MLP) for various flow configurations. The results showed that both GNN and MLP models yield reasonable prediction overall. However, GNN shows superior performance near-wall due to spatial convolution. Although the present method implements the rotational invariant discretely, the augmentation method is found to produce consistent performance for any rotated coordinates. The minimal flow configuration, which can train a model to predict a range of flow configurations, is also explored. It is found that a model trained based on turbulent channel flows alone yields a close level of prediction robustness to the ones trained with multiple flow configurations. The developed GNN model is implemented in OpenFOAM, and large eddy simulation (LES) results are compared with corresponding direct numerical simulation data. With these proposed techniques, ML-based SGS models can be improved in terms of robustness and usability for a range of LES applications.
Many underlying relationships among data in several areas of science and engineering, e.g., computer vision, molecular chemistry, molecular biology, pattern recognition, and data mining, can be represented in terms of graphs. In this paper, we propose a new neural network model, called graph neural network (GNN) model, that extends existing neural network methods for processing the data represented in graph domains. This GNN model, which can directly process most of the practically useful types of graphs, e.g., acyclic, cyclic, directed, and undirected, implements a function tau(G,n) isin IRm that maps a graph G and one of its nodes n into an m-dimensional Euclidean space. A supervised learning algorithm is derived to estimate the parameters of the proposed GNN model. The computational cost of the proposed algorithm is also considered. Some experimental results are shown to validate the proposed learning algorithm, and to demonstrate its generalization capabilities.
Stable density stratification introduces buoyancy effects, vortex–wave coupling, and internal-wave radiation into body wakes, making full-field prediction more challenging than in homogeneous flows. This study develops a residual-gated mesh graph neural network (RGGNN) for coupled density–velocity prediction of two-dimensional stratified circular-cylinder wakes using large-eddy-simulation data. The model combines a mesh-based encoder–processor–decoder architecture with residual and gated incremental updates to reduce autoregressive error accumulation and preserve local flow structures. It is evaluated through single-condition autoregressive rollout, cross-Froude-number prediction under linear stratification, and cross-condition prediction under hyperbolic-tangent stratification, together with a direct-transfer test on an unseen tandem-cylinder wake. At the 10th rollout step, the RGGNN reduces the root mean square errors of density, horizontal velocity, and vertical velocity by 96.5%, 84.4%, and 86.1%, respectively, relative to the baseline GNN. The extended 49-step rollout further shows substantially slower error growth. For the unseen-Froude-number case, the density root mean square error decreases from 0.099 37 to 0.005 12, while the coefficient of determination increases from 0.989 29 to 0.999 97. Under hyperbolic-tangent stratification, the density root mean square error is reduced by 76.3%–83.1%. In the unseen tandem-cylinder test, the RGGNN maintains coefficients of determination above 0.994 for all predicted variables without retraining. These results demonstrate improved rollout stability, local-feature preservation, and preliminary transferability across unseen geometry and mesh topology, indicating the potential of residual-gated mesh graph learning for coupled density–velocity prediction of stratified wakes within controlled numerical settings.
This study introduces a novel physics-informed spatiotemporal graph neural network (PIST-GNN) for rapid, physically consistent, and accurate long-term prediction of complex water entry flow fields. The model integrates a MeshGraphNet-based GNN for spatial feature extraction on unstructured meshes and a long short-term memory (LSTM) network for temporal evolution, critically enhanced by embedding mass conservation and volume of fluid (VOF) transport equations as soft constraints during training. Comprehensive evaluations demonstrate that the combined physics constraints provide an optimal balance, achieving robust predictions with only a slight increase in data loss while significantly reducing physical inconsistencies like velocity divergence and VOF advection residuals. The PIST-GNN exhibits superior long-term stability and generalization capabilities, consistently maintaining the lowest prediction errors and the best physical consistency over 30 autoregressive steps on unseen Froude number and initial angle cases. Furthermore, through both quantitative and qualitative analysis of prediction results, we elucidate the complex influence of initial conditions on prediction accuracy. Ultimately, this study provides a practical and feasible approach to address the persistent challenge of error accumulation in long-term predictions of water entry flow fields by purely data-driven models. Future research will focus on overcoming the limitations of the current two-dimensional slice approximation method by optimizing physical constraints and model architecture, thereby more directly capturing the complete three-dimensional dynamic characteristics of the water entry flow field.
Graph and convolutional neural network coupling with a high-performance large-eddy simulation solver
No abstract available
Abstract A neural-network-based large eddy simulation is performed for flow over a circular cylinder. To predict the subgrid-scale (SGS) stresses, we train two fully connected neural network (FCNN) architectures with and without fusing information from two separate single-frame networks (FU and nFU, respectively), where the input variable is either the strain rate (SR) or the velocity gradient (VG). As the input variables, only the grid-filtered variables are considered for the SGS models of G-SR and G-VG, and both the grid- and test-filtered variables are considered for the SGS models of T-SR and T-VG. The training data are the filtered direct numerical simulation (fDNS) data at $Re_d=3900$ based on the free-stream velocity and cylinder diameter. Using the same grid resolution as that of the training data, the performances of G-SR and G-VG (grid-filtered inputs) and T-SR-FU and T-VG-FU (grid- and test-filtered inputs with fusion) are better than those of the dynamic Smagorinsky model and T-SR-nFU and T-VG-nFU (grid- and test-filtered inputs without fusion). These FCNN-based SGS models are applied to untrained flows having different grid resolutions from that of training data. Although the performances of G-SR and G-VG are degraded, T-SR-FU and T-VG-FU still provide good performances. Finally, T-SR-FU and T-VG-FU trained at $Re_d = 3900$ are applied to higher-Reynolds-number flows ($Re_d = 5000$ and 10 000) and their results are also in good agreements with those of fDNS and previous experiment, indicating that adding the test-filtered variables and fusion increases the prediction capability even for untrained Reynolds number flows.
Abstract As the Reynolds number increases, the large-eddy simulation (LES) of complex flows becomes increasingly intractable because near-wall turbulent structures become increasingly small. Wall modeling reduces the computational requirements of LES by enabling the use of coarser cells at the walls. This paper presents a machine-learning methodology to develop data-driven wall-shear-stress models that can directly operate, a posteriori, on the unstructured grid of the simulation. The model architecture is based on graph neural networks. The model is trained on a database which includes fully developed boundary layers, adverse pressure gradients, separated boundary layers, and laminar–turbulent transition. The relevance of the trained model is verified a posteriori for the simulation of a channel flow, a backward-facing step and a linear blade cascade.
Abstract In this paper, we propose artificial-neural-network-based (ANN-based) nonlinear algebraic models for the large-eddy simulation (LES) of compressible wall-bounded turbulence. An innovative modification is applied to the invariants and the tensor bases of the nonlinear algebraic models through using the local grid widths along each direction to normalise the corresponding gradients of the flow variables. Furthermore, the dimensionless model coefficients are determined by the ANN method. The modified ANN-based nonlinear algebraic model (MANA model) has much higher correlation coefficients and much lower relative errors than the dynamic Smagorinsky model (DSM), Vreman model and wall-adapting local eddy-viscosity model in the a priori test. The significantly more accurate estimations of the mean subgrid-scale (SGS) fluxes of the kinetic energy and temperature variance are also obtained by the MANA models in the a priori test. Furthermore, in the a posteriori test, the MANA model can give much more accurate predictions of the flow statistics and the mean SGS fluxes of the kinetic energy and the temperature variance than other traditional eddy-viscosity models in compressible turbulent channel flows with untrained Reynolds numbers, Mach numbers and grid resolutions. The MANA model has a better performance in predicting the flow statistics in supersonic turbulent boundary layer. The MANA model can well predict both direct and inverse transfer of the kinetic energy and temperature variance, which overcomes the inherent shortcoming that the traditional eddy-viscosity models cannot predict the inverse energy transfer. Moreover, the MANA model is computationally more efficient than the DSM.
Abstract A fully connected neural network (NN) is used to develop a subgrid-scale (SGS) model mapping the relation between the SGS stresses and filtered flow variables in a turbulent channel flow at $Re_\tau = 178$. A priori and a posteriori tests are performed to investigate its prediction performance. In a priori test, an NN-based SGS model with the input filtered strain rate or velocity gradient tensor at multiple points provides highest correlation coefficients between the predicted and true SGS stresses, and reasonably predicts the backscatter. However, this model provides unstable solution in a posteriori test, unless a special treatment such as backscatter clipping is used. On the other hand, an NN-based SGS model with the input filtered strain rate tensor at single point shows an excellent prediction capability for the mean velocity and Reynolds shear stress in a posteriori test, although it gives low correlation coefficients between the true and predicted SGS stresses in a priori test. This NN-based SGS model trained at $Re_\tau = 178$ is applied to a turbulent channel flow at $Re_\tau = 723$ using the same grid resolution in wall units, providing fairly good agreements of the solutions with the filtered direct numerical simulation (DNS) data. When the grid resolution in wall units is different from that of trained data, this NN-based SGS model does not perform well. This is overcome by training an NN with the datasets having two filters whose sizes are bigger and smaller than the grid size in large eddy simulation (LES). Finally, the limitations of NN-based LES to complex flow are discussed.
Abstract An artificial neural network (ANN) is used to establish the relation between the resolved-scale flow field and the subgrid-scale (SGS) stress tensor, to develop a new SGS model for large-eddy simulation (LES) of isotropic turbulent flows. The data required for training and testing of the ANN are provided by performing filtering operations on the flow fields from direct numerical simulations (DNSs) of isotropic turbulent flows. We use the velocity gradient tensor together with filter width as input features and the SGS stress tensor as the output labels for training the ANN. In the a priori test of the trained ANN model, the SGS stress tensors obtained from the ANN model and the DNS data are compared by computing the correlation coefficient and the relative error of the energy transfer rate. The correlation coefficients are mostly larger than 0.9, and the ANN model can accurately predict the energy transfer rate at different Reynolds numbers and filter widths, showing significant improvement over the conventional models, for example the gradient model, the Smagorinsky model and its dynamic version. A real LES using the trained ANN model is performed as the a posteriori validation. The energy spectrum computed by the improved ANN model is compared with several SGS models. The Lagrangian statistics of fluid particle pairs obtained from the improved ANN model almost approach those from the filtered DNS, better than the results from the Smagorinsky model and dynamic Smagorinsky model.
No abstract available
A subgrid-scale (SGS) model for large-eddy simulation (LES) of compressible isotropic turbulence is constructed by using a data-driven framework. An artificial neural network (ANN) based on local stencil geometry is employed to predict the unclosed SGS terms. The input features are based on the first-order and second-order derivatives of filtered velocity and temperature which appear in the second-order Taylor approximation of the SGS stress and heat flux. It is shown that the proposed ANN-7 model performs better than the gradient model in the a priori test. The correlation coefficient is larger and the relative error is smaller for ANN-7 model as compared to those of the gradient model in the a priori test. In an a posteriori analysis, the performance of ANN-7 model shows advantage over the dynamic Smagorinsky model and dynamic mixed model in the prediction of spectra and structure functions of velocity and temperature, and instantaneous flow structures. Artificial neural network is a promising tool for understanding the physical fundamentals of SGS unclosed terms with further improvement.
Artificial neural network mixed model for large eddy simulation of compressible isotropic turbulence
No abstract available
Turbulence closure modeling remains one of the central challenges in computational fluid dynamics. Over the past decade, machine learning (ML) has emerged as a promising paradigm to augment and, in some cases, replace classical turbulence models by leveraging high-fidelity simulation data and advanced neural network architectures. We argue that a single obstacle recurs across the field—a mismatch between the pointwise, offline objective on which data-driven closures are trained and the solver-coupled, multi-scale, statistically stationary way they must perform—and we trace it to intrinsic physical properties of turbulence (extreme multi-scale chaos, non-locality, and the absence of scale separation). Through this lens, we survey three established pillars: (i) ML-enhanced Reynolds-averaged Navier–Stokes (RANS) closures, (ii) ML-based subgrid-scale models for large-eddy simulation, and (iii) physics-informed approaches, notably physics-informed neural networks (PINNs). Their characteristic failure modes—RANS ill-conditioning, the a priori/a posteriori gap, and PINN spectral bias—are local signatures of this common root. Emerging paradigms—neural operators, graph neural networks, transformers, diffusion models, and foundation models—are assessed by whether they change the objective or merely the architecture, and we propose concrete protocols for testing whether learned fields respect the Navier–Stokes dynamics statistically rather than pointwise. We critically examine generalizability across flow configurations, interpretability, numerical stability under solver coupling, and the scarcity of high-fidelity data, and conclude that aligning data-driven methods with physical constraints—and the training objective with deployment—is the path toward reliable, generalizable, and computationally efficient turbulence models for engineering applications.
No abstract available
<p>In this work, we consider turbulence closures of LES (Large Eddy Simulation) type for classical decaying 2D turbulence in a priori and a posteriori experiments using explicit filtering approach. According to Germano 1986 decomposition, full subfilter stress for Gaussian filter is decomposed into three Galilean-invariant parts: Leonard, Cross and Reynolds stresses. By analysing spectral transfer of energy and enstrophy, we show that Leonard stress redistributes resolved energy toward large scales and dissipates substantial part of enstrophy, while Cross stress provides an additional enstrophy dissipation at subfilter scales and Reynolds stress predominantly injects energy into middle scales (i.e., Kinetic Energy Backscatter). Substantial part of enstrophy dissipation is located on subfilter scales, and it should be accounted for by choosing base filter wide enough compared to mesh step of LES model. Otherwise, significant fraction of enstrophy dissipation will correspond to subgrid scale stress, which is less universal and harder to approximate. As a result of a priori analysis, we propose LES closure consisting of three parts: SSM (Scale Similarity Model), which is equivalent to Leonard stress, biharmonic Smagorinsky damping as a Cross stress counterpart and ADM (Approximate Deconvolution Model) approximation for Reynolds stress ("backscatter"). The proposed model have two free parameters: Smagorinsky constant and amplitude of the backscatter. These parameters are estimated in a posteriori experiments utilizing dynamic approach and energy-enstrophy balance equation, correspondingly. The proposed model have the following distinctive features: it reproduces energy and enstrophy transfer spectra in accordance to the individual components of the subfilter forces, "reproduces" base filter and reproduces energy growth in accordance to the filtered DNS (Direct Numerical Simulation) solution.</p><p>The work was supported by the Russian Foundation for Basic Research (projects 19-35-90023, 18-05-60184) and Moscow Center for Fundamental and Applied Mathematics (agreement with the Ministry of Education and Science of the Russian Federation No. 075-15-2019-1624).</p>
We present a fundamentally new autonomic subgridscale closure for large eddy simulations (LES) that solves a nonlinear, nonparametric system identification problem instead of using a predefined turbulence model. The autonomic approach expresses the local SGS stress tensor as the most general unknown nonlinear function of the resolvedscale primitive variables at all locations and times using a Volterra series. This series is analogous to a Taylor series expansion in both time and space, and incorporates nonlinear, nonlocal, and nonequilibrium turbulence effects. The series introduces a large number of convolution kernel coefficients that are found by solving an inverse problem to minimize the error in representing known subgrid-scale stresses at a test filter scale. The optimized coefficients are then projected to the LES scale by invoking scale similarity in the inertial range and applying appropriate renormalizations. This new closure approach avoids the need to specify a turbulent constitutive model and instead identifies an optimal model on the fly. Here we present the most general formulation of the new autonomic approach and outline an inverse modeling method for optimizing the coefficients. We then explore truncations of the series expansion and demonstrate the effects of regularization and sampling on the optimal coefficients. Finally, we perform a priori tests of this approach using data from direct numerical simulations of homogeneous isotropic and sheared turbulence. We find substantial improvements over the Dynamic Smagorinsky model, even for a 2nd order time-local truncation of the present closure.
No abstract available
The analysis of kinetic energy spectrum and dissipation mechanisms plays a crucial role in both theoretical examination and real-world implementations within compressible turbulent flows. The previous works within the near-dissipation range are conducted within the framework of incompressible turbulent flow dynamics. We investigate the kinetic energy spectrum and transfer within the near-dissipation range of compressible isotropic turbulent flows via highly resolved direct numerical simulations with different turbulent Mach numbers (Mt). We found that in compressible turbulent flows, the kinetic energy spectrum model is significantly influenced by the turbulent Mach number, with a more pronounced tail as the turbulent Mach number increases. Within the near-dissipation range, there exist strong local spatial fluctuations of the kinetic energy flux with the increase in grid resolution. The viscous dissipation can be decomposed into solenoidal and dilatational components, and its dilatational component leads to stronger intermittency within the near-dissipation range. Besides, the high turbulent Mach number also leads to a stronger dissipation, which serves as an alternative factor for the high intermittency within the near-dissipation range in compressible turbulent flows. The uncovered high intermittency provides a theoretical foundation for the prediction of intense events and the development of turbulent modelings.
No abstract available
No abstract available
A novel method, known as the vector potential random flow generation (VPRFG) method, is introduced for synthesizing divergence-free homogeneous isotropic turbulence with arbitrary spectra. First, the proposed approach employs the random-flow-generation-based method to create a vector potential field. Subsequent application of the curl operation to this field produces a turbulent flow that inherently satisfies the divergence-free condition. In the formulas of the proposed method, we explicitly impose arbitrary homogeneous isotropic three-dimensional spatial cross-spectral density (CSD) and Taylor's frozen hypothesis. This ensures that the generated turbulence conforms to prescribed statistical characteristics, including energy spectra, one-dimensional spatial power spectral density (PSD), temporal PSD, spatial coherence function, turbulent kinetic energy, and Reynolds stress. Additionally, the theoretical accuracy of the proposed method is validated through numerical examples, employing the von Kármán energy spectrum as the target value. Finally, large eddy simulations of homogeneous isotropic temporal-decaying box turbulence generated by the VPRFG method demonstrate a close alignment with the experimental results.
In the present work, we begin with assessing a well-known spectral closure in large eddy simulations (LES) of isotropic turbulence at high Reynolds numbers. The original model has been developed assuming the existence of an infinite inertial range energy spectrum. The results show that although the energy is transferred among resolved flow scales at approximately correct rate, some characteristic features of the inertial subrange are not recovered despite the existence of the k−5/3 scaling. The issue appears to follow from insufficient resolution of flow scales relevant to LES and suboptimal functional form of the spectral eddy-viscosity, leading to excessive energy dissipation in the low-wavenumber range. Using a differential regulator which requires a predefined energy spectrum, we develop a new model that adaptively determines the spectral eddy-viscosity, taking into account the finite spatial resolution of the simulation. The results confirm that the adaptively determined closure facilitates the development of the inertial subrange and has a negligible computational overhead. Moreover, we demonstrate that the LES enriched with the newly developed model can provide reliable insight into the physics of large flow scales, particularly at high Reynolds numbers. We introduce a new resolution criterion for LES which is expressed in terms of the Taylor microscale λ and estimate that a significant part of the inertial subrange in isotropic turbulence at Reλ∼10 000 can be accurately resolved with at most ∼16 3843 modes, which is currently achievable using modern exascale computing resources.
Abstract Accurate prediction of turbulent flows remains a critical challenge in computational fluid dynamics, particularly for complex, non-equilibrium environments where traditional Reynolds-averaged Navier–Stokes (RANS) models fail to capture deep anisotropy and nonlocal interactions. While data-driven turbulence modeling, such as tensor basis neural networks (TBNNs) offers a promising alternative, these approaches frequently struggle with physical realisability and the inherent nonlocality of turbulent transport. To address these limitations, this work introduces a realisability-preserving E(3)-equivariant graph neural network (E(3)-GNN) for Reynolds-stress anisotropy closure. By leveraging equivariant message passing and a spherical-tensor formulation, the architecture explicitly captures nonlocal neighborhood interactions while guaranteeing Galilean invariance. Furthermore, a stabilized positive semi-definite projection ensures that all predicted states permanently reside within the physically admissible Lumley triangle, completely eliminating the unphysical predictions that traditionally destabilize coupled solvers. Because accurate a priori stress predictions do not guarantee a posteriori stability during active RANS simulation, we additionally propose a solver-in-the-loop fine-tuning methodology using Group Relative Policy Optimization (GRPO). This step directly optimizes a scalar production-correction policy within the active RANS environment to capture dynamic energy propagation. Evaluated on the parameterised periodic hill benchmark family, with training and testing performed on distinct hill geometries, our E(3)-GNN achieves strict physical realisability by construction. Through GRPO fine-tuning, the fully coupled model successfully recovers the intense turbulent kinetic energy (k) peaks and mean velocity profiles in the separated shear layer that purely supervised models fail to capture, bridging the gap between offline regression accuracy and stable, coupled solver performance.
A predictive model has been developed based on experimental data obtained from open channel flow past two horizontal cylinders, one above the other, using a graph neural network framework. Upon investigating the error plots, it is evident that the decreasing mean square error indicates the model's capability to predict near accurate data. Furthermore, second- and third-order structure functions have been illustrated to validate Kolmogorov's 2/3 and 4/5 laws for both original and predicted data, which were found to be a little inconsistent near the two cylinders in the near-wake zone. To investigate this inconsistency further, scaling exponents from extended self-similarity have been calculated and plotted with the conventional Kolmogorov 1941 theory in order to visualize the deviation. It is found that strong anisotropic flow occurs in the vicinity of the rough bed and two cylinders, which resulted in substantial intermittency. This article bears a significant contribution to the understanding of turbulence intermittency and deep learning implementation in turbulent open channel flow.
Intracranial aneurysms (IAs) are life-threatening vascular conditions requiring accurate risk assessment to guide treatment. Hemodynamic biomarkers such as wall shear stress and oscillatory shear index are promising predictors of rupture risk but remain underused clinically due to the high computational cost of traditional CFD methods. We propose a physics-constrained graph neural network (GNN) framework trained on high-fidelity CFD data to predict full 3D, time-resolved hemodynamic fields throughout the cardiac cycle. Our model incorporates enhanced node features and physics-based constraints to capture complex spatio-temporal flow behavior in near real time. It generalizes to varying inflow conditions and unseen patient-specific geometries with no fine-tuning. Additionally, we release a benchmark dataset of 105 patient-derived aneurysm geometries with CFD fields to support the machine learning (ML) community. This is the first GNN model applied to transient 3D aneurysmal flow prediction, paving the way for rapid, AI-driven hemodynamic analysis toward risk stratification and treatment planning.
In Computational Fluid Dynamics (CFD), approximate solutions of the Navier-Stokes equations are obtained using discretization methods on computational grids. The design of such grids has a significant influence on the accuracy of the approximate solution. While a refined grid improves the accuracy in general, this directly leads to increased computational cost. It is best practice to conduct grid convergence studies to quantify the discretization error on coarser grids and to find the optimal trade-off between cost and accuracy. However, obtaining grid converged results for full aircraft configurations at flight Reynolds numbers is often regarded as infeasible in an industrial setting, where it is not uncommon to use grids with several hundred million points. Under these conditions, it is often impossible to conduct rigorous grid convergence studies.
Accurate prediction of urban wind flow is essential for urban planning and environmental assessment. Classical computational fluid dynamics (CFD) methods are computationally expensive, while machine learning approaches often lack explainability and generalizability. To address the limitations of both approaches, we propose Diff-KNN, a hybrid method that combines Coarse-Scale CFD simulations with a K-Nearest Neighbors (KNN) model trained on the residuals between coarse- and fine-scale CFD results. Diff-KNN reduces velocity prediction errors by up to 83.5% compared to Pure-KNN and 56.6% compared to coarse CFD alone. Tested on the AIJE urban dataset, Diff-KNN effectively corrects flow inaccuracies near buildings and within narrow street canyons, where traditional methods struggle. This study demonstrates how residual learning can bridge physics-based and data-driven modeling for accurate and interpretable fine-scale urban wind prediction.
This paper explores several methods intended to predict corrections to improve coarse grid solutions of a flow field. Motivated by applications in hemodynamics, this flow is on the transition to turbulence, and simulations are performed on unstructured tetrahedral grids. The predictions are based on learning deep neural networks with differences of coarse and fine grid solutions. Two kinds of feedforward networks, several graph neural networks, and a hierarchical neural network are studied. To simplify the task, the correction of the statistically steady state is considered. However, even then, all studied approaches fail to produce useful results in the validation step, which reveals limitations of current deep learning approaches to predict coarse grid corrections for complicated flow problems.
Unsteady aerodynamic phenomena, such as gusts, turbulence, and fluid-structure interactions affect an aircraft during flight. For design, optimisation and certification, it is indispensable to quantify such unsteady aerodynamic effects. Industry-standard computational fluid dynamics methods, such as solving the unsteady Reynolds-averaged Navier-Stokes equations or the linearized frequency domain method, are either computationally expensive or restricted by assumptions like linearity. Once trained, machine learning methods are capable of computing non-linear relationships very fast, making them suitable as surrogate models. By autoregressively applying graph neural networks (GNNs), operating on a discretised spatial domain, spatio-temporal predictions can be made. However, autoregressive GNNs suffer from error accumulation leading to unstable rollouts over time. Here we show that combining GNNs with augmented Neural Ordinary Differential Equations yields temporally stable predictions of the surface forces on a pitching airfoil. We found that our approach, called GNODE, based on Graph Neural Ordinary Differential Equations, provides temporally more stable, spatially smoother, and overall more accurate results than an autoregressive GNN baseline. Tests are conducted on a dataset consisting of a simulations of a pitching airfoil, including transonic shocks, transient behaviour and dynamic non-linearities. Augmenting GNODEs with additional latent dimensions improves the expressivity and accuracy by capturing underlying history effects. The developed method demonstrates an approach that is suitable to model non-linear spatio-temporal systems with exogenous inputs.
Dental aerosol procedures produce sub-50 micrometre nuclei that can remain airborne for long periods in enclosed clinics, creating pathways for airborne pathogen transmission. Reynolds-Averaged Navier-Stokes (RANS) simulations with Euler-Lagrange particle tracking capture this transport accurately but require very long run times per scenario, which precludes real-time clinical decision support in 3D. We present the Eulerian-Lagrangian Graph Interaction Network (ELGIN), a physics-informed graph surrogate that jointly predicts carrier-flow dynamics on the OpenFOAM polyhedral mesh and the per-parcel motion of the polydisperse spray cloud. ELGIN couples a multi-head Graph Transformer with Jacobi-preconditioned learnable pressure projection and a turbulence-closure head to a sigmoid-gated Lagrangian Interaction Network through differentiable inverse-distance mesh-parcel coupling, and advances parcels with a symplectic Stormer-Verlet integrator. A four-stage physics-informed curriculum stabilises 260-step autoregressive rollouts without gradient explosion. A parameter sweep with foam-extend 4.1 OpenFOAM reactingParcelFoam across clinically relevant ventilation rates and handpiece spray speeds provides CFD ground truth. This article reports a single-case demonstration in which both ELGIN and a Lagrangian-only baseline (M0) are trained and evaluated on Sweep_Case_03 of a twenty-case sweep; full 16/2/2 retraining is in progress and will replace all reported metrics. On this case, ELGIN tracks the foam-extend particle cloud much more closely than M0: mean parcel displacement error falls from 19.56% to 16.20% of room width and cloud radius-of-gyration error from 9.85% to 6.58%. A 26-second rollout completes in ~64 s on a 4 GB GPU, approximately 37x faster than the foam-extend reference pipeline, toward per-appointment infection-risk screening once the multi-case checkpoint is in place.
This work proposes a novel methodology for turbulence modeling in Large Eddy Simulation (LES) based on Graph Neural Networks (GNNs), which embeds the discrete rotational, reflectional and translational symmetries of the Navier-Stokes equations into the model architecture. In addition, suitable invariant input and output spaces are derived that allow the GNN models to be embedded seamlessly into the LES framework to obtain a symmetry-preserving simulation setup. The suitability of the proposed approach is investigated for two canonical test cases: Homogeneous Isotropic Turbulence (HIT) and turbulent channel flow. For both cases, GNN models are trained successfully in actual simulations using Reinforcement Learning (RL) to ensure that the models are consistent with the underlying LES formulation and discretization. It is demonstrated for the HIT case that the resulting GNN-based LES scheme recovers rotational and reflectional equivariance up to machine precision in actual simulations. At the same time, the stability and accuracy remain on par with non-symmetry-preserving machine learning models that fail to obey these properties. The same modeling strategy translates well to turbulent channel flow, where the GNN model successfully learns the more complex flow physics and is able to recover the turbulent statistics and Reynolds stresses. It is shown that the GNN model learns a zonal modeling strategy with distinct behaviors in the near-wall and outer regions. The proposed approach thus demonstrates the potential of GNNs for turbulence modeling, especially in the context of LES and RL.
Large-eddy simulations (LES) require closures for filtered production rates because the resolved fields do not contain all correlations that govern chemical source terms. We develop a graph neural network (GNN) that predicts filtered species production rates on non-uniform meshes from inputs of filtered mass fractions and temperature. Direct numerical simulations of turbulent premixed hydrogen-methane jet flames with hydrogen fractions of 10%, 50%, and 80% provide the dataset. All fields are Favre filtered with the filter width matched to the operating mesh, and learning is performed on subdomain graphs constructed from mesh-point connectivity. A compact set of reactants, intermediates, and products is used, and their filtered production rates form the targets. The model is trained on 10% and 80% blends and evaluated on the unseen 50% blend to test cross-composition generalization. The GNN is compared against an unclosed reference that evaluates rates at the filtered state, and a convolutional neural network baseline that requires remeshing. Across in-distribution and out-of-distribution cases, the GNN yields lower errors and closer statistical agreement with the reference data. Furthermore, the model demonstrates robust generalization across varying filter widths without retraining, maintaining bounded errors at coarser spatial resolutions. A backward facing step configuration further confirms prediction efficacy on a practically relevant geometry. These results highlight the capability of GNNs as robust data-driven closure models for LES on complex meshes.
State-of-the-art deep learning models have been extensively utilized to reconstruct small-scale structures from coarse-grained data in turbulent flows. However, their application has predominantly been restricted to structured uniform meshes, limiting their applicability to data associated with complex geometries that are typically simulated on structured non-uniform or unstructured meshes. Machine learning (ML) models based on graph neural networks (GNNs), known for their ability to process unstructured data, offer a promising alternative. In this study, we leverage the inherent flexibility of GNNs featuring message passing layers to develop a methodology for reconstructing unresolved small-scale structures from low-resolution data on complex meshes. The accuracy of the proposed approach is demonstrated using two cases: a reacting channel flow on a structured non-uniform mesh, and a reacting hydrogen fueled internal combustion (IC) engine featuring an unstructured mesh. Evaluation of results based on visual agreement, statistical metrics, and cumulative error reduction indicates the effectiveness of the method in accurately reconstructing fine-scale features. Overall, this study provides a pathway for integrating data-driven small-scale reconstruction and subgrid-scale modeling to enhance the accuracy of coarse-grained simulations on complex meshes.
We propose an arbitrary Lagrangian-Eulerian (ALE)-consistent machine learning framework for long-term fluid-structure interaction (FSI) prediction on deforming unstructured meshes. Specifically, the fluid dynamics are modeled by a surrogate that combines a graph neural operator (GNO) with a vision Transformer (ViT) for spatiotemporal prediction, while a lightweight long short-term memory (LSTM) network predicts structural kinematics at the interface. The two surrogates are coupled through a standard partitioned procedure. Most importantly, kinematic compatibility at the moving interface is enforced via an ALE-consistent boundary-correction step that updates the fluid-side interface velocity with the predicted structural velocity at each coupling update, thereby improving near-interface accuracy and long-term rollout stability. To mitigate autoregressive error accumulation, a two-stage training strategy is adopted, consisting of single-step supervised pretraining followed by long-term autoregressive fine-tuning. The proposed framework is validated on the benchmark problem of a flexible beam vibration in the wake of a cylinder. Results demonstrate accurate phase-consistent predictions over long rollouts and robust generalization under inlet-profile variations in both interpolation and extrapolation settings. Systematic ablation studies further assess the respective contributions of the ViT module, ALE-consistent boundary correction, and long-term training to predictive accuracy and rollout robustness.
The differentiable physics paradigm may be leveraged as an a-posteriori approach for discovering turbulence closure models by embedding a neural network parameterization directly inside the solver and optimizing it given potentially sparse target data. This addresses a key limitation of a-priori learning where direct numerical simulation (DNS) data is used to approximate the subgrid stress with the assumption of a low-pass filter. Closures trained in this a-priori manner frequently lead to unstable deployments due to the mismatch between the assumed filter and the effect of numerical discretizations and coarse-graining. In comparison, while typically stable during deployment, a-posteriori learning incurs high computational costs due to the need to backpropagate through a large eddy simulation (LES) solver. Furthermore, a-posteriori methods are challenging to apply broadly since they require significant modification of existing solvers. Finally, both approaches are limited when generalization is desired across different numerical schemes with their implicit filtering characteristics. In this work, we present a deep-learning approach for turbulence closure modeling built on the continuous data assimilation framework. Our approach enables the a-priori training of closures using sparsely observed DNS data without modifying or differentiating through the LES solver, while preserving stability during deployment for the recovery of invariant statistics. We focus on the model's ability to adapt to different discretizations by explicitly conditioning it on the numerical scheme. We use two- and three-dimensional canonical cases to test our framework and show that the learned correction systematically tracks the discretization error of the coarse solver.